read-memories

Overview of read-memories skill

What read-memories does

read-memories is a Claude Code skill for searching prior session logs so you can recover decisions, repeated patterns, open loops, and user corrections. It is most useful when someone asks “do you remember,” “what did we decide,” or when a new task clearly depends on earlier context. If you need the read-memories skill for Workflow Automation, think of it as a retrieval step that turns old chat history into usable working context.

Who should install it

Install read-memories if you work across multiple Claude Code sessions, maintain ongoing projects, or frequently revisit unfinished work. It is a good fit for agents and power users who need continuity more than fresh ideation. It is less useful if your tasks are mostly isolated, one-shot requests with no dependency on previous conversations.

What makes it different

The key differentiator is that the skill is not just a generic prompt about memory; it is a concrete log-search workflow with a silent query step and a follow-up synthesis step. That means the main value is not “remembering” in the abstract, but quickly finding evidence in local session files and using it without exposing raw logs to the user.

How to Use read-memories skill

Install and scope it correctly

Use the read-memories install flow by adding the skill to your Claude Code setup, then keep the skill available only in environments where local session logs exist. The skill expects access to Bash and a standard Claude Code project log layout. If you plan to use read-memories for Workflow Automation, make sure the runtime can read the log paths it queries.

Start from the right input

The skill works best when you pass a specific keyword or topic that likely appears in prior logs, such as a feature name, ticket ID, customer name, or decision phrase. A weak request like “check memory” is too broad. A stronger prompt looks like: “Use read-memories to find prior decisions about the DuckDB export flow and summarize any unresolved issues.” That gives the skill a clear search target and an output goal.

What the skill actually runs

The repository’s workflow is simple: query session logs with DuckDB, then internalize the results before responding. The read-memories usage pattern is built around searching JSONL files under $HOME/.claude/projects/ and optionally limiting the search to the current project with --here. If you are adapting the skill, read SKILL.md first, then inspect the exact search path logic and the two-step handling of results.

Tips that improve results

Use distinct keywords, not vague concepts. Add project context when the same term appears in many places. When the first search returns too much noise, narrow by project or by a more specific phrase. If the output is meant to inform a current task, tell the skill what kind of memory matters most: decisions, TODOs, user preferences, or prior mistakes. That changes what it should extract from the logs.

read-memories skill FAQ

Is read-memories only for Claude Code?

It is designed around Claude Code session logs and the local project structure shown in the repo. That makes it a strong fit for that ecosystem, but not a universal memory tool. If your workflow does not store conversations in the expected JSONL paths, the skill will not be a good match.

How is it different from a normal prompt?

A normal prompt asks the model to infer context. read-memories uses actual log search first, then summarizes the evidence. That means it is better when correctness depends on prior decisions, not when you just want a fresh answer. The read-memories guide is therefore about retrieval, not general reasoning.

Is it beginner-friendly?

Yes, if you can name what you are trying to recover. The hard part is not using Bash; it is choosing a keyword that likely appears in past sessions. If you are unsure, start with a project name plus a task label, then refine after the first search.

When should I not use it?

Do not use read-memories when there are no relevant local logs, when the topic is truly new, or when the request depends on external facts rather than prior sessions. In those cases, a standard research or drafting workflow is faster and cleaner.

How to Improve read-memories skill

Give it better search terms

The biggest quality lever is the keyword. Instead of a broad noun, use names, short phrases, or identifiers that were likely repeated in the original discussion. For example, “refund policy” is weaker than “Q4 refund exception” or a ticket number. Better inputs produce better recall and less irrelevant log noise.

Ask for the right kind of memory

Be explicit about the output you want from read-memories usage: decisions, unresolved questions, preferences, blockers, or next steps. If you only ask for “context,” the result can be too diffuse. If you ask for “find the last agreed API contract and any objections,” the skill can filter more intelligently.

Watch for common failure modes

The main failure modes are overly broad searches, duplicate log hits, and over-trusting a single old session. A useful read-memories skill review should check whether the retrieved context is still current and whether later sessions contradicted it. If the first pass surfaces stale information, rerun with a narrower project scope or a newer keyword.

Iterate after the first result

Use the first retrieval to identify better terms, then search again with those terms if needed. This is especially helpful when a topic evolved over time and the initial query only finds the earliest mention. For read-memories for Workflow Automation, the best practice is to treat the first pass as discovery, then refine the search before you rely on the memory in an automated workflow.